How to Get the Most from Claude Opus 5.5
Original: Getting the most out of Opus 5.5 in Claude and Claude Code
Why This Matters
Longer autonomous runs shift how developers delegate multi-step engineering work to AI agents.
Anthropic's Addy Osmani published a practical guide on Sept 22, 2026 for using Claude Opus 5.5 in Claude apps and Claude Code, covering prompting, steering long runs, and checking results.
Opus 5.5 runs longer autonomously than its predecessor and thinks before every reply—no need to add 'think carefully' to prompts. Anthropic says removing such lines actually made replies start sooner in chat testing with no quality drop. The model's biggest improvement over Opus 5 is on multi-step, multi-file work: early testers ran coding tasks for hours with minimal oversight.
The guide's core advice: give the whole task in one message, name a concrete finish line ('tests pass,' 'old client deleted'), and specify only when the model should stop to ask. Mid-run additions are possible in Claude Code—type a follow-up and hit Enter while the model works, avoiding costly restarts.
For design outputs, vague instructions like 'avoid a generic look' mostly swap one default style for another. The guide recommends listing specific patterns to exclude—cream backgrounds, monospace labels, pill buttons—then iterating on whatever the model picks instead.
The model controls its own thinking depth; users can adjust effort in Claude Code but cannot otherwise override it.